Download E-Customization - Semantic Scholar

ASIM ANSARI and CARL F. MELA*
Customized communications have the potential to reduce information
overload and aid customer decisions, and the highly relevant products
that result from customization can form the cornerstone of enduring customer relationships. In spite of such potential benefits, few models exist
in the marketing literature to exploit the Internet’s unique ability to design
communications or marketing programs at the individual level. The
authors develop a statistical and optimization approach for customization
of information on the Internet. The authors use clickstream data from
users at one of the top ten most trafficked Web sites to estimate the
model and optimize the design and content of such communications for
each user. The authors apply the model to the context of permissionbased e-mail marketing, in which the objective is to customize the design
and content of the e-mail to increase Web site traffic. The analysis suggests that the content-targeting approach can potentially increase the
expected number of click-throughs by 62%.
E-Customization
Marketers have long realized the value of targeting and
customization. Customized products and communications
attract customer attention and foster customer loyalty and
lock-in. Targeted communications aid customer decisions
and reduce information overload, and highly relevant products yield satisfied customers. The customer loyalty that
results from such personalization and targeting can translate
into increased cash inflows and enhanced profitability. Customized marketing solutions are useful for both customer
acquisition and retention and can engender successful, longterm relationships. However, customization has often
proved difficult because of implementation challenges,
insufficient customer information, and other factors.
The Web makes mass customization eminently possible.
Firms can exploit the capabilities afforded by digitization
and networking to provide unique content of direct relevance to each customer. Moreover, such tailoring of information can be done quickly and at low cost. Customization
is possible in part because of the interactivity afforded by
the Web. Firms can collect and update preference information of customers from on-site surveys and from the traces
customers leave as they navigate through a Web site. This
knowledge can then be seamlessly integrated with algorithms and software to customize content automatically for
individual consumers. Indeed, customized design (serving
different variants of content to different users at different
points in time) represents one of the key features that differentiate the Web from more traditional media.
In the online world, content sites such as C-net and Yahoo
can leverage a loyal customer base to increase readership
and therefore advertising revenues.1 Various surveys have
extolled the rapid growth in advertising dollars on the Web.
Given the large magnitude of expected revenue involved,
content providers are increasingly turning toward customization strategies to increase their share of advertising
income. Similarly, e-commerce sites such as Amazon.com
and Dell can customize content (e.g., information, digital
products such as software, advertising, promotions, and
other incentives) to increase repeat purchases and crossselling.2
In spite of this potential, few models exist to help firms
implement one-to-one marketing on the Internet. Therefore,
given the substantial potential arising from e-customization,
it is our objective to develop a statistical and optimization
approach for customization of information on the Internet. A
secondary goal is to model clicking behavior on the Internet.
Our procedures enable firms to customize both the content
(what and how much information) and the design (rendition)
*Asim Ansari is an associate professor, Graduate School of Business,
Columbia Business School (e-mail: [email protected]). Carl F. Mela is
Associate Professor of Marketing, The Fuqua School of Business, Duke
University (e-mail: [email protected]). The authors are listed alphabetically.
The authors thank Bill Boulding, Andrew Gershoff, Don Lehmann, Kamel
Jedidi, John Lynch, Richard Staelin, Seenu Srinivasan, and the seminar participants at Columbia and Duke Universities for their helpful comments.
Special thanks to Rajeev Kohli for many insightful discussions and help on
the optimization. Jeffrey Inman served as guest editor for this article.
1Lehman Brothers (Becker, Anmuth, and Leichter 2002) forecasts that
online advertising will grow with a 12% compound annual growth rate
from $5.3 billion in 2002 to $7.7 billion in 2005.
2The online market now exceeds $171 billion and is growing rapidly.
131
Journal of Marketing Research
Vol. XL (May 2003), 131–145
132
to their users. Specifically, we develop an approach that
enables Web sites to customize permission-based e-mail
communications to increase Web site traffic (though the
approach can be more broadly applied to the issue of Web
site customization). As such, we contribute to the growing
literature on one-to-one marketing (e.g., Rossi, McCullouch, and Allenby 1996; Shaffer and Zhang 1995).
However, e-customization differs from some other contexts inasmuch as (1) the “product” (the e-mail or Web site
itself) is complex and requires explicit optimization for customization; (2) the product can be constructed dynamically,
thereby making customization truly possible; and (3) the
media are highly addressable, thereby facilitating targeting.
Moreover, although it is possible to target products or marketing tactics in other contexts, the costs of doing so (e.g.,
printing individual catalogs, the cost of point-of-sale
coupons) can often be exceedingly high.
E-customization requires knowledge of individual-level
preferences, and it is therefore important to accommodate
unobserved sources of preference heterogeneity (by allowing model parameters to vary across consumers). Moreover,
the efficacy of content and design is likely to differ across
implementations. Although it is possible to identify and
account for some of the factors affecting the response to
content and design in a typical application, it is unlikely that
a Web marketer can enumerate all the variables that affect
consumer response. Accordingly, our approach also models
sources of unobserved contextual heterogeneity across email content and design. This enables Web marketers to better predict potential responses to a particular type of content
on a particular e-mail (or, in the case of an on-site targeting
strategy, a particular type of content on a particular Web
page design).
In the marketing literature, sources of heterogeneity are
typically modeled by means of a random-coefficients
approach. The computer science literature on Web customization suggests the use of collaborative filtering
approaches to model heterogeneity (Breese, Heckerman,
and Kadie 1998). Collaborative filtering systems use data
from users with similar preferences to recommend new
items. Such systems form the basis for most commercial
recommendation systems (e.g., Netperceptions, Macromedia). We show how model-based collaborative filtering can
be implemented using a Bayesian semiparametric model.
Specifically, we show how a mixture of Dirichlet process
(MDP) probit model can be used to perform collaborative
filtering and to flexibly accommodate different sources of
unobserved heterogeneity. The Bayesian literature in marketing has predominantly used normal distribution to capture differences across customers (Allenby and Rossi 1999;
Ansari, Essegaier, and Kohli 2000; Rossi, McCulloch, and
Allenby 1996). The normal distribution has limited flexibility as it is unimodal, has thin tails, and does not accommodate skewness. These sources of inflexibility of the normal
distribution could result in misleading inferences and inaccurate individual-level estimates (Escobar 1994). The MDP
model, in contrast, is flexible enough to accommodate deviations from normality and, depending on the date, can automatically adjust to mimic either a finite mixture of support
points or a continuous distribution for heterogeneity,
whichever is appropriate. Moreover, because of its discrete
JOURNAL OF MARKETING RESEARCH, MAY 2003
representation of heterogeneity, it can mimic a collaborative
filtering representation. In this article, we explicitly compare
results from a MDP specification with those obtained from
models that use normal distributions to capture
heterogeneity.
Given a set of parameter estimates, customization then
requires the construction of customized e-mails for each
consumer. We therefore develop an optimization procedure
for customized e-mail design. The optimization procedure
uses as input the individual-level parameter estimates from
the hierarchical Bayesian statistical model and enables firms
to (1) select relevant information to include in an e-mail and
(2) configure the content to enhance the probability of site
visits. Although the permutations in design can be quite
large, we provide an exact solution to the design problem.
Our design problem shows how the promise of targeting that
is offered by hierarchical Bayesian models can be brought to
fruition.
The rest of the article proceeds as follows: In the next section, we detail various approaches to customization on the
Internet. We then describe our statistical model for clicking
behavior and provide the details of our application. Next, we
overview the data, specify the model, discuss parameter estimates, and perform model comparisons. Then, we present
the e-mail design optimization problem and the optimization
approach for obtaining optimal configurations. Finally, we
offer conclusions and suggest future research directions.
CUSTOMIZATION APPROACHES
Web sites can use a combination of on-site and external
customization approaches to manage customer relationships. Both approaches are useful in enhancing site loyalty
because they increase switching costs for users. When users
are faced with a decision to switch sites, it is quite feasible
that they will be reticent to invest the time to begin “training” another firm. As Alba and colleagues (1997) note, consumers might expect to experience switching costs and a
decrease in customer service were they to switch to another
site.
On-Site Customization
In this approach, companies either customize the Web site
to appeal to users or enable the users themselves to customize the content. For example, portal sites such as
Netscape and Altavista enable users to self-customize the
site. Users of such sites can specify keywords of interest to
filter news stories, can provide lists of stocks for which they
require regular information, or can manipulate the page
views themselves. Such user-initiated customization has
obvious advantages, as it elicits user preferences and gives
control to users in defining what they want.
However, in many cases, such a user-initiated approach
may not be completely successful, as users may not be able
to fully or accurately self-explicate their preferences. Many
novice users may not feel confident about performing such
customization actions. Moreover, preferences are dynamic
and users may be reluctant to provide information continually (or may find it cumbersome to do so). In such situations,
company-initiated customization based on revealed preferences data may be more useful.
E-Customization
On many Web sites, company-initiated on-site customization occurs in the form of recommendation systems. For
example, firms such as Amazon.com and Kraft use recommendation systems to suggest products (Ansari, Essegaier,
and Kohli 2000; Gershoff and West 1998) or content to customers.3 Most commercial recommendation systems (e.g.,
Netperceptions, Macromedia) use techniques such as collaborative or content filtering on customer ratings data to
determine customers’ product preferences. Recommendations can also be made with attribute-based approaches. For
example, Ansari, Essegaier, and Kohli (2000) show that
hierarchical Bayesian models are best suited for recommendation systems because they incorporate different sources of
heterogeneity and provide individual-level estimates even in
sparse data environments.
These recommendation systems have typically been oriented toward suggesting a new product (e.g., a movie) or
service rather than designing Web pages or e-mails.
Although these approaches could be adapted to the customization of content (by extending them to consider multiple, concurrent recommendations), they are more difficult to
adapt to issues of customized design, because design is a
large-scale (many control variables) optimization problem
(e.g., how to design a new product as opposed to recommending an existing one). Our approach therefore generalizes these previous works.
External Customization
An external customization approach is intended to bring
users to a Web site. Typically, e-mails, banner advertisements, affiliate sites, or other communication media herald
site content that may be of interest to site users. For example, companies such as Amazon.com, Morningstar, and The
New York Times regularly send e-mails containing hypertext
links to the content of their Web sites to registered users. Emails intended to attract customers to a site typically contain
(1) brief summaries of editorial content and (2) a link (or a
set of links) to the site, where more detailed information can
be found. After reading the summaries, users can click on
the link listed in the e-mail. By learning user preferences
from clicking histories and demographics, Web sites can tailor the messages in the e-mail to the user’s interests. The
greater the history of users’ information, the more likely a
firm can learn (and thus match) users’ interests.
E-mail is one of the most popular Internet applications
and is rapidly being adopted for e-commerce. Between 1999
and 2000, e-mail revenues increased 270% to $342 million
and are expected to grow to $1 billion by 2003 (Aberdeen
Group 2001). As click-through rates on banner advertisements continue to drop, e-mail is becoming the instrument
of choice for business-to-consumer communication. Customization of e-mails to suit the preferences of users is
3In addition to content recommendations, some companies provide navigational aids to help customers interact with the information provided on
the Web sites. Perkowitz and Etzioni (1997) show how Web sites can use
log-file analysis to suggest index pages of links to users. Such link prediction can also be used to fetch documents in advance while the user is reading a page. Alternatively, Web sites can suggest navigational shortcuts predicated on most popular navigational patterns. Sarukkai (2000) describes
such a tour generation procedure based on Markov chains.
133
therefore of paramount importance. Many companies, such
as Doubleclick, Clickaction.com, Netperceptions, and
Macromedia, have developed e-mail marketing systems to
assist companies in outbound e-mail marketing. The specific
details of their implementations, unfortunately, are not publicly available.
We focus on an external customization application for
several reasons (our particular application involves personalizing permission-based e-mail design and content to
attract the e-mail recipients to a Web site). First, external
customization is relatively easy to implement, as firms do
not need to create a significant number of alternative Web
site designs (according to a Gartner Group study [Satterthwaite 1999], an average site costs $1 million and five
months of implementation time). Second, internal customization strategies rely on the visitor to come to the site,
which can take some time. In contrast, external customization strategies can be effective immediately, as communications are sent directly to the user by e-mail, post, or advertising. Direct communications, such as e-mails, enable firms
to entice users more actively to the site and are therefore
useful for both acquisition and retention activities. Third,
on-site customization of the Web site can be risky if users
have become familiar with the interface; changes to the
home page can confuse loyal users. Fourth, as noted previously, e-mail targeting is an important and growing application in its own right.
Note that the application of our algorithm to e-mail does
not preclude its use on Web page customization. Indeed,
often e-mails are sent as Web pages that are opened directly
by a Web browser. Yet porting our analysis to the redesign of
Web sites would require careful deliberation. First, dynamic
Web sites may confuse users, which suggests that greater
benefit may accrue to varying content than design. Second,
successful Web site customization is incumbent on reliably
identifying site visitors and may thus be of limited use when
dynamic Internet protocol (IP) addresses manifest or cookies are disabled. Third, as e-mails are served daily (or less
frequently), there is ample time to estimate models and
serve new content between points of contact with the user.
Applications such as ours are not likely to scale well to onthe-fly customization, and more research is needed in that
area. Nonetheless, it is possible to update Web sites on a
daily or weekly basis, and substantial benefits might obtain
even at this frequency of customization.
MODELING APPROACH
Using parameter estimates to custom design e-mails, we
develop an individual-level model for estimating the probability of clicking on links within e-mails. Given that our
objective is to customize content and design, our modeling
approach consists of two phases:
Phase 1: In the first phase, we specify and estimate a probability model that correlates content and design characteristics to individual clicking likelihoods. The input to
this model is individual-level clickstream data of past
responses to content links included in e-mails. The
output is a probability function and a set of individuallevel and e-mail–level parameters that represent the
preference structures of the users and the differences
across e-mails.
134
JOURNAL OF MARKETING RESEARCH, MAY 2003
Phase 2: In the second phase, we use the probability model and
the individual-level preferences as input to an optimization model. The optimization model recommends
the optimal e-mail configuration (content and layout)
for each recipient on each occasion.
As the optimization model is predicated on the results
obtained from the econometric model, we first describe our
statistical model and its results.
In the statistical model, we use an attribute-based
approach and model the customer responses in terms of email design attributes, content descriptors, and user characteristics. The database contains click-through responses of
users for content links delivered in different e-mails. Let i =
1 to I index users, j = 1 to J represent e-mails, and k = 1 to
K indicate the distinct links for which user response data are
available. Each customer i provides binary responses yijk for
ni2 links over ni1 e-mails. Let Ei = {j1, j2, …, jni1} be the
index set of e-mails sent to user i and let Li = {k1, k2, …,
kni1} be the index set of links for which user i’s responses
are available. Users differ in the number of observations
(links), and similarly e-mails differ in the number of links
they contain, thus yielding a highly unbalanced data set. The
total number of observations in the data set is given by N =
Σni2.
The observed binary responses (clicks), yijk, can be modeled through a random-utility framework. Users click on a
particular link when the utility for exploring the content
associated with the link exceeds a threshold. The relation
between the observed response and the latent utility of clicking can be written as
(1)
0 if u ijk ≤ 0.
y ijk = 
1 if u ijk > 0
We model the latent utility uijkfor link k of e-mail j for
user i as the function of observed and unobserved e-mail and
link characteristics. Specifically, the utility function can be
written as
(2)
u ijk = x ′ijk µ + z ′jk λ i + w ′ik θ j + γ k + e ijk ,
where eijk ~ N(0, 1). The vector xijk contains observed
user-, e-mail- (design variables), and link-level (content)
variables. The coefficients in µ contain the “fixed effects”
and describe the population-level impacts of the independent variables. The remaining terms specify the random
effects and are used for capturing different sources of heterogeneity. Note that the error specification in Equation 2
assumes that errors are independent across the different
links. Although it is possible to relax this assumption, we
refrain from doing so for several reasons. First, such a solution would not be scalable, especially when the number of
content categories is large. The introduction of dependencies through a multivariate probit specification leads to a
substantial increase in computational complexity of the estimation and optimization algorithms, making broader implementation of our model difficult. Second, the introduction of
covariance terms between the utilities of links makes customized design difficult, as customization would require the
knowledge of the correlations between each new link and all
others, as well as the evaluation of a multivariate integral.
Third, our model predicts response well even though we
assume independence (we predict 649 e-mail clicks in our
data and observe 639). Therefore, in our view, the costs of
this approach outweigh the benefits.
It is unlikely that all factors that affect responses can be
isolated in any given application. It is therefore crucial to
allow for multiple sources of heterogeneity. For example,
users may differ in their preferences for content and in their
propensities to click on links in different portions of the email. Moreover, these differences in preferences may be
unrelated to user demographics and other observed user
variables. To allow for such unobserved preference heterogeneity across users, we introduce the term z′jkλi in the utility function. The vector zjk can contain a subset of the variables in xijk such as the content descriptors and design
variables pertaining to the link k and the e-mail j. The
individual-specific coefficients in λi then indicate how the
content and design preferences for individual i differ from
the population average, µ.
In addition to modeling preference heterogeneity, in
applications involving responses to information products, it
is also important to model contextual heterogeneity. Emails, being information products, are complex entities and
therefore are not completely amenable to simplistic featurebased renditions. For example, some e-mails may be better
designed than others, and the design features may interact in
intricate patterns, which makes it difficult to code an e-mail
on the basis of few observable attributes. Thus, e-mails may
differ in terms of both observed and unobserved attributes.
It is therefore desirable to account for the contextual influences with a random-effects approach. Accordingly, we use
the term w′ikθj to capture the differential impact of e-mail j
on the utility function. The vector w′ik can contain link-level
and individual-level variables. The coefficients in θj indicate
how e-mail j differs from the average e-mail in terms of the
impact of link-level and individual-level variables on clickthrough.
A broad categorization of the content of a link is possible
based on a few features. However, it is likely that the content remains only partially explained in terms of the
observed variables. For example, although a news item can
be broadly classified as a “business news” item, there can
still be considerable variation in content among all business
news items. We therefore use a random effect γk to accommodate unobserved content heterogeneity.
Dirichlet Process Priors
In this section, we show how semiparametric distributional assumptions on the population distribution of the random effects can yield a principled approach for model-based
collaborative filtering. The random effects are assumed to
come from a population distribution. The marketing literature has used either finite mixture distributions (Wedel and
Kamakura 2000) or continuous distributions (Ansari, Jedidi,
and Jagpal 2000) to represent population heterogeneity.
Finite mixtures allow flexibility, but in complex models, it is
difficult to determine the appropriate number of components. Moreover, it is not straightforward to incorporate
multiple sources of heterogeneity in finite mixture models.
Alternatively, continuous distributions are used as part of
hierarchical Bayesian models to capture heterogeneity. In
this case, typically a normal population distribution is used
to represent the variation in random effects. Although the
choice of the normal distribution is made for tractability and
E-Customization
conjugacy reasons, this assumption may not necessarily
hold in reality. The normal distribution provides limited
flexibility because it is unimodal, has thin tails, and does not
accommodate skewness. If the population distribution is not
normal, misleading inferences about the magnitude of
effects and the nature of heterogeneity are possible.
Researchers have used finite mixtures of normal components (Allenby, Arora, and Ginter 1998) to circumvent these
problems, but in such models the difficulty of determining
the number of components remains.
In this article, we show how an MDP model (Escobar
1994; MacEachern 1994) can be used to model heterogeneity in a flexible yet structured manner. The MDP model
enables us to capture the uncertainty about the functional
form of the population distribution using a semiparametric
approach. This model avoids the typical assumption of a
parametric population distribution such as the normal and
instead uses an unknown distribution F to model heterogeneity. As this population distribution is assumed to be
random, in the MDP model, a Dirichlet process prior
(Blackwell and MacQueen 1973; Ferguson 1973, 1974) is
placed on the population distribution F. The Dirichlet
process provides a mechanism of placing a probability distribution on the space of distributions.
The Dirichlet process prior F ~ D(F|F0, α) is described by
two parameters: F0 is a parametric baseline distribution that
defines the “location” of the Dirichlet process prior, and α is
a positive scalar precision parameter that determines the
concentration of the prior for F about the baseline distribution F0. The baseline distribution F0 can be considered a
prior “guess” for the population distribution. In this article,
we use a normal distribution as the baseline distribution. The
precision parameter α determines how close the nonparametric distribution F is to the baseline distribution F0. When
α is large, a randomly sampled population distribution F is
very similar to F0. Therefore, if the baseline distribution F0
is normal and the precision parameter α → ∞, then the population distribution is a discrete distribution that mimics a
normal distribution. In contrast, when α is small (α → 0),
the sampled population distribution has its mass concentrated on a few points and is therefore similar to a finite mixture distribution.4
As the precision parameter is inferred from the data, the
MDP specification allows flexible incorporation of heterogeneity. If the nature of the heterogeneity is consistent with
a normal distribution, the precision parameter is automatically adjusted to be large and the MDP yields a population
distribution that mimics a normal. Conversely, if the data
come from a nonnormal population distribution, the MDP
model allows enough flexibility because of its semiparametric nature and, as do finite mixture models, accommodates
deviations from normality. In addition, the number of “segments” is automatically determined by the MDP algorithm
as outlined subsequently. Thus, the MDP model places
fewer restrictions on the shape of the population distribution, and as the population distribution is well approximated, it can have beneficial consequences for the accuracy
of individual-level estimates (see Escobar 1994).
135
In the context of our model, we assume that the user-, email-, and link-specific random effects come from population distributions that arise from different Dirichlet process
priors. Specifically, we use
(3)
to model the user-specific random effects. Equation 3
assumes that the user-specific random effects come from an
unknown population distribution F1. The population distribution in turn comes from a Dirichlet process prior with a
multivariate normal baseline distribution that has a mean 0
and an unknown covariance matrix Λ. The precision parameter of the Dirichlet process prior, α1, controls how close the
sampled population distribution is to the baseline normal
distribution.
Similarly, the e-mail random effects can be modeled as
( 4)
(1994) indicates how sample size interacts with α.
θ j ~ F2
F2 ~ D[ N(0, Θ), α 2 ],
where the baseline distribution is a normal with Θ as the
covariance matrix. Finally, the link-specific random effects
can be modeled as
(5)
γ k ~ F3
F3 ~ D[ N(0, τ), α 3 ],
where τ represents the variance of the associated univariate
baseline distribution.
A Bayesian approach is needed for inference regarding
the unknown parameters of the MDP model. The unknown
θj},{γk}, Λ,
λi}, {θ
quantities in our model include {{u}, µ, {λ
, , α1, α2, α3}. The joint posterior distribution cannot be
written in closed form, and therefore Markov chain Monte
Carlo (MCMC) methods (see Bush and MacEachern 1996;
Doss 1994; West, Muller, and Escobar 1994) are needed to
sample from the posterior distribution. The MCMC methods
involve sampling iteratively from the full-conditional
distributions.
To understand further how the MDP model implements
model-based collaborative filtering and to explicate how the
MDP model differs from the popular approach of using normal population distributions, we contrast here the fullconditional distribution of the user-specific random effects
λi obtained from these models. The full-conditional
expresses the uncertainty about λi given the values of the
other unknowns. In contrasting these full-conditional distributions, let ũ ijkλ = u ijk − x ′ijk µ − w ′ik θ j − γ k represent the
adjusted utility for an observation. Then ũ ijkλ ~ N(z′jkλi, l).
Form the vector ũ i by stacking the adjusted utilities ũ ijkλ for
all the observations of the user, and form the matrix Zi by
stacking row by row all the row vectors z′jk for the observations belonging to user i. When λi is assumed to be distributed normal N(0, Λ) (as is the case for one of our null models), it is well known that the full-conditional distribution for
the random effects λi is multivariate normal and can be written as
(6)
4Escobar
λ i ~ F1
F1 ~ D[ F1 | N(0, Λ ), α1 ]
p(λ i |{u ijk}, µ, {θ j}, {γ k}, Λ ) ~ N(λˆ i , Vi ),
where the posterior precision is given by Vi−1 = Λ−1 + Z ′i Z i,
and the posterior mean is given by λˆ i = Vi + Z ′i u˜ i .
136
JOURNAL OF MARKETING RESEARCH, MAY 2003
In contrast, for the MDP model, the full conditional for λi
is given by the following mixture of a normal distribution
and a mass point distribution:
( 7)
p(λ i |{λ k , k ≠ i}, {u ijk}, µ, {θ j}, {γ k}, Λ )
~ q p0 Fb (λ i |.) +
∑q
l≠i
pl δ λ l ,
where
λi|.) is the baseline posterior distribution given in
•Fb(λ
Equation 6.
•The weight associated with the normal component is
qp0 ∝ α1fi, where fi is the marginal density of the
adjusted utilities for user i under the multivariate normal
baseline prior density N(0, Λ). The marginal density is
obtained by integrating out the user random effects—that
λi,µ
µ,{θ
θj},{γk})N(0, Λ)dλ
λi) and is multivariate
is, ∫f(ũi|λ
normal N(0, ZiΛZ′i + Ini), where Ini is the identity matrix.
The quantity fi is obtained by evaluating at ũi, the marginal density.
λl,µ
µ,{θ
θj},{γk}), the normal density of the utili•qpl ∝ f(ũi|λ
ties for user i evaluated by means of user l’s parameters;
that is, each qpl is proportional to the multivariate normal
density of ũi ~ N(Ziλl,Ini).
The weights qpk are standardized to sum to 1. The parameter δs indicates a degenerate distribution with point mass at
s. Therefore, in Equation 7, with probability proportional to
qp0 we sample λ i from the full-conditional under the baseline population distribution; that is, using Equation 6 and
probability proportional to qpl, we select from the degenerate distribution δλl, which means that we set λi = λ1 (i.e., we
set person i’s parameters to be the same as person l’s). This
results in a mixture with one component being a normal distribution, and all other components are point masses on the
parameters of other users.
Intuitively, this mixing scheme implies that in any iteration of the MCMC scheme, if the likelihood of observing
user i’s data is relatively large using user l’s parameters, then
the random effect λ1 for person l is more likely to be chosen
as user i’s random effect. In this case, users i and l would be
in the same cluster. Conversely, if the likelihood of observing i’s data is relatively low when user l’s random effect is
used for user i, the more likely it is that the i’s random effect
is a new value (generated from the baseline distribution or
from a different user, l′, whose parameters generate a higher
likelihood for i’s data). This mixing scheme results in a clustering of the random effects because users share common
random-effect parameters on any given iteration. However,
the number of “clusters” and the allocation of users to the
clusters change from one iteration to another. Thus, this
mechanism for generating the user-specific random effects
is similar to what can be called model-based collaborative
filtering on parameter space, as information from similar
users is used to predict the users’ preferences. In contrast to
the standard model, in which the estimates for a user depend
on the data for other users only through the population mean
and variance, here, the posterior of λi heavily depends on the
data for the user and the data of “nearest neighbors.”
APPLICATION
Data
The data for this project are furnished by one of the Web’s
leading Internet sites. As a condition for their use, the sponsoring firm wishes to remain anonymous, and any identifying aspects of the data are therefore disguised. The organization derives the majority of its revenue from selling
advertising space on its Web site to client firms and is therefore highly interested in increasing site usage. To accomplish that, it sends e-mails to registered users inviting them
to visit the site. These e-mails include a synopsis of several
articles on the site. Below each article summary is a link to
the article (the link is a Web address that readers can click
on to go to the site). It is the response to these links that
forms our dependent measure.
To clarify our exposition of the site and its content further,
we depict the site as an automotive news and reviews site
(though the content is not automotive). We also categorize
the site’s content (including the links it sends in its e-mails)
much like an automotive site can be categorized as cars or
trucks. These categorizations are denoted content areas and
are established by the sponsoring firm on the basis of its
knowledge of the industry. On this site, a particular content
area (e.g., the car portion of the site) sends a daily e-mail to
users who register to receive the e-mail. These e-mail links
pertain to content types within the content area. For example, the link types for a car area may include car reviews, car
pricing, car specifications, automotive news, and so forth.
Upon receiving this e-mail, users might click on a particular
type of link. If they do, this information is recorded. There
are two key components to the data set provided by this
firm: (1) the user log files that record usage history for a
given respondent and (2) the e-mail files that provide the
date, content, and design of the e-mails.
User log files. Each time the visitor to the site clicks on an
e-mail or Web site link, a record is generated and stored. The
record contains the time, the origin of the click (the IP
address), and information about the link or Web page that
was clicked. These records, collected between June and
August 1999, form one of the two key portions of the data
we analyze. The origin of the click (the clicker’s identity) is
determined by “cookies,” or records placed on the visitor’s
(clicker’s) hard drive, as well as the user’s IP address. When
the user visits the site again, the content provider makes note
of the cookie on the user’s computer to ascertain whether he
or she visited the site before.
E-mail files. Visitors to the content provider’s site can
request to receive e-mails pertaining to information on the
site. Only people who are registered receive e-mails; therefore, all recipients had registered.5 In addition to the e-mail
addresses of the users, the e-mail files contain the dates of
the e-mails, the links listed on the e-mails, and a synopsis of
information contained on those links. This information is
used to code the design and content of the links provided in
the e-mail and to infer which links were not clicked. The
5Note that our model applies equally well to contexts in which users are
not registered. In this case, e-mail addresses are obtained from the purchase
of e-mail lists from third-party vendors.
E-Customization
design of the e-mail includes (1) the amount (number) of
links listed in the e-mail, (2) the order of the links, and (3)
the e-mail type (html or text). The link content in our data is
coded into content categories provided by the firm. For
example, an article reviewing an automobile would be coded
“review,” and automotive news would be coded “news.” All
in all, there are 12 of these content categories. Finally, information regarding the links in the e-mails can be merged with
the database on users’ clicking histories to determine which
links (if any) the user clicked.
Sample size. All totaled, there are three months of e-mail
file data and two months of log file data. The number of emails averaged about five per week, and there are 1048 users
in our sample.6 The number of links per e-mail averaged
approximately 5.6 and ranged from 2 to 8. The average
response rate across links is approximately 7%. We used
data from a random sample of 100 users to estimate the
models and assigned 60% of the e-mails (11,436 observations) into an estimation sample and 40% of the e-mails into
validation sets.
We created two such validation data sets. In the first, we
randomly held out observations across all e-mails, users, and
links, so the validation set included extant e-mails and extant
links. Therefore, predictions regarding a user’s likelihood of
clicking on a given link could be made with information
regarding how others clicked on that link. The size of this
validation set, which we denote “Extant,” was 3735 observations. In the second validation data set, the sample was
composed entirely of new links and e-mails. Thus, only
information regarding users’ past behavior could be used to
forecast the likelihood of clicking on a link. This validation
data set, which we denote “Novel,” consisted of 3900
observations.
Ansari, Essegaier, and Kohli (2000) outline the benefits of
creating multiple validation data sets. In contexts in which
extant communications are to be targeted to additional consumers who have yet to receive them, the Extant data set is
more relevant. Such may be the case after a test e-mail or
when an e-mail is sent to additional customers of a site. In
contexts in which new content is to be targeted, the Novel
data set may be more relevant. Such may be the case when
new editorial content is generated, as in the case of daily
news.
Model Specification
Previously, we developed a general random-effects model
that can be applied across a number of contexts. Here we
describe the specific instantiation for our application. We
include several observed link variables, e-mail design variables, and user variables in our specification. In addition to
these observed effects, we allow for different sources of
unobserved heterogeneity across users, e-mails, and links.
Our specification can be described as follows:
6Only users who responded to at least one e-mail per month are included
in the sample. People who never or rarely respond represent 24% of the
sample but only 2% of the clicks. Because of their lack of response, there
is little information available to improve this number. Therefore, customization is likely to be inefficacious for these people, so we omit them
from our analysis. We note that the practice of including only respondents
has an analog in scanner data research pertaining to category incidence
(Manchanda, Ansari, and Gupta 1999; Mela, Jedidi, and Bowman 1998).
137
(8) u ijk = µ1 + µ 2 Content1 + L + µ12 Content11 + µ13Position
+ µ14 Num-Items + µ15Text + µ16Since + λ i1
+ λ i 2 Content1 + L + λ i,12 Content11 + λ i,13Position
+ λ i,14 Num-Items + λ i,15Since + θ j1 + θ j2 Content1
+ L + θ j,12 Content11 + θ j,13Position + θ j,14Since
+ γ k + e ijk .
Link variables. The link variables characterize both the
editorial content of the link and its position within the email. The likelihood of clicking on a link will be a function
of its content (e.g., news versus reviews) because users are
expected to differ in their preference for content categories.
The editorial content within the e-mails can be categorized
into 12 categories (or content types). These content types are
included by means of a set of 11 dummy variables (Content
1–Content 11).
As is the case with traditional media, the placement of the
link may affect response (click-through). Hanssens and
Weitz (1980) find that the later the advertisement is presented in a magazine, the less likely it is to be seen or read.
Hoque and Lhose (1999) have replicated this result in electronic media. They argue that the impact of placement is
magnified in electronic media because it is more difficult to
read online and because of the effort involved in scrolling.
The ordinal position of the link within the e-mail is represented by a variable (Position).
E-mail variables. We specify two e-mail design variables.
The first represents the number of links within the e-mail
(Num-Items). Houston and Scott (1984) show that increasing the number of advertisements within a print environment
lowers the readership of any particular advertisement on a
page. We expect a similar result in electronic settings as an
increase in the number of links exacerbates clutter and therefore increases the cognitive costs of perusing the e-mail. For
some users, the cost of needing to scan a large number of
links may exceed the potential value of reading the e-mail.
The second e-mail–level variable (Text) characterizes
whether the e-mail is a text e-mail or an html e-mail.7
Person variable. As the duration of time increases since
the last link clicked, the likelihood of a subsequent click
may change. Therefore, we include a covariate for days
since the last click (Since). We use one observation to initialize this variable. In customer relationship management
applications, an increase in time since the last purchase
leads to a decrease in subsequent purchase likelihood
(Schmittlein, Morrison, and Colombo 1987). Therefore, we
expect that the coefficient for Since will typically be negative, as users who have not clicked in some time are less
likely to be active. The descriptive information for the link,
e-mail, and person variables is presented in Table 1.
User heterogeneity. We model unobserved user heterogeneity by specifying a random-effects model for both the
intercepts and the slopes in our model. Thus, we specify (1)
user-specific random intercepts (which capture differences
across users in their propensities to click on links) and (2)
7We tested for an interaction between Num-Items and Position. This did
not enhance model fit. We also tested for nonlinear effects for these variables and found none.
138
JOURNAL OF MARKETING RESEARCH, MAY 2003
Table 1
VARIABLE DESCRIPTORS
Calibration Data
Variable
Mean
Standard
Deviation
Link Click
E-mail Click
Text
Num-Items
Position
Since
Content1
Content2
Content3
Content4
Content5
Content6
Content7
Content8
Content9
Content10
Content11
.075
.307
.635
5.609
3.305
7.197
.181
.182
.182
.069
.062
.061
.036
.036
.022
.037
.032
.263
.461
.481
.784
1.656
9.372
.385
.386
.386
.255
.242
.239
.187
.187
.148
.188
.177
user-specific random slopes (which capture differences
across users in their responses to link content, e-mail variables, and time since last click). We note that user and random effects are identified only for variables that are not
fixed across users (similarly, e-mail random effects are identified only for variables that are not fixed across e-mails).
E-mail heterogeneity. We capture e-mail heterogeneity by
allowing (1) an e-mail–specific random intercept that captures e-mail “attractiveness” and (2) e-mail–specific random
slopes that capture differences across e-mails. The unobserved e-mail variables (captured by θj in Equation 8) interact with the link-level variables such as the content of the
link and the position of the link within the e-mail. These
interactions enable us to model contextual heterogeneity. As
it is not possible to completely describe e-mails using
observed attributes, incorporating e-mail–level heterogeneity is crucial for modeling click-through probabilities.8
Link heterogeneity. For parsimony, we capture link heterogeneity by including a single link-specific random effect.
This term captures the impact of unobserved link effects—
for example, the particular editorial (news or review) content of a link type on a given day—that is left unexplained
by the content variables describing the link.
RESULTS
We estimated three models on the data. The first model is
a simple model (Model S) that includes no heterogeneity.
The second model (Model N) includes all observed variables and accounts for user-, e-mail-, and link-specific
unobserved sources of heterogeneity, using normal population distributions for the random effects. The third model
(Model DP) uses Dirichlet process priors to account for
user, e-mail, and link heterogeneity.
We estimated the models using MCMC methods. The
full-conditional distributions used in MCMC sampling for
the DP model are available on request from the authors. For
8We
thank the guest editor for noting that e-mail heterogeneity may also
capture potential confounds between content and position.
Extant Validation Data
Novel Validation Data
Mean
Standard
Deviation
Mean
Standard
Deviation
.077
.278
.623
5.620
3.310
8.111
.180
.182
.182
.065
.064
.064
.038
.037
.023
.036
.035
.250
.448
.485
.803
1.663
10.204
.384
.386
.386
.246
.244
.245
.191
.189
.150
.187
.184
.087
.350
.621
5.784
3.392
6.160
.179
.179
.179
.068
.065
.052
.035
.033
.033
.036
.031
.282
.477
.485
1.021
1.747
8.011
.384
.384
.384
.252
.246
.221
.184
.179
.178
.185
.173
each model, we ran the chain for 30,000 iterations. The
results reported are based on a sample of 20,000 draws from
the posterior distribution, after we discarded 10,000 burn-in
draws. We ensured convergence by monitoring the time
series of draws.
Model Selection
Model fit. We use the pseudo-Bayes factor (PsBF)
(Geisser and Eddy 1979; Gelfand 1996) to compare different models. The PsBF is based on the cross-validation predictive density, which we can conveniently compute for our
models using the MCMC draws. Let y be the observed data,
yil represent the lth observation for user i, and y(il) represent
the data with the observation l for user i deleted. The crossvalidation predictive density can be written as
( 9)
π( y il |y( il ) ) =
∫ π(y |β, y
il
( il )
) π(β|y(il) )dβ,
where β is the vector of all parameters in the model. The
PsBF for comparing two models (M1 and M2) is expressed
in terms of the product of cross-validation predictive densities and can be written as
ni
I
(10)
PsBF =
∏∏
i =1
l =1
π[( y il )| y ( il ) , M1]
π[( y il )|y ( il ) , M 2]
.
The PsBF summarizes the evidence provided by the data for
M1 against M2, and its value can be interpreted as the number of times model M1 is more (or less) probable than model
M2.
The PsBF for our model can be calculated easily from a
β1, …, βd}. As β is the vector
sample of d MCMC draws {β
of all parameters, including all the random effects, the
binary responses yil, i = 1 to I, l = 1 to ni, are conditionally
independent given β. Thus, a Monte Carlo estimate of
π(yil|y(il)) can be obtained as
(11)
πˆ ( y il | y( il ) )

1
= 
d

d
∑
t =1
−1

 ,
(p(ilt ) )y il (1 − p(ilt ) )1 − y il 
1
E-Customization
where p(t)
probability Pr(yil = 1|β(t)) = 1 – (u(t)
il is the
il –
(t)
′
(t)
′
′ θ (t) – γ (t) , where the superscript t
xilµ – zj(l)k(l)λ i – wik(l)
j(l)
k(l)
represents the tth draw of the MCMC sampler and j(l)
denotes the e-mail associated with the lth observation and
k(l) represents the link associated with that observation.
Gelfand (1996) provides the derivation for Equation 11 for
the cross-validation predictive distribution. The estimates
from Equation 11 can be used to calculate the logarithms of
the numerator and denominator of the PsBF. These quantities can be considered a surrogate for the log-marginal data
likelihoods for the two competing models. On the basis of
the MCMC output, the log-marginal data likelihood for
Model DP is –2290.25, for the normal model (Model N) is
–2340.45, and for the nonheterogeneous model is –2807.22.
Accordingly, the PsBF implies an exp(50.2) improvement
for Model DP over Model N and an exp(517.0) improvement for Model DP over Model S. Thus, the PsBFs provide
support for the Dirichlet process model.
Model Predictions
The predictive ability of the three competing models can
be assessed through the validation data. The link-level probabilities for the links within an e-mail can be used to compute an e-mail–level probability of at least one click-through
from the e-mail. To compute this probability, we first use the
observed data to determine the likelihood of clicking on
each link in the e-mail, p̂ ijk. The complement of these predictions yields the likelihood that the link was not clicked,
1 − p̂ ijk. Under the assumption of independence, the product
of these link-level nonclick likelihoods yields the joint probability that the e-mail was not clicked. The complement of
ˆ
this probability, pˆ ij = [1 − Π K
k = 1 (1 − p ijk )], yields the probability that at least one link was clicked. The independence
assumption seems plausible, as it predicts 649 e-mail clicks
compared with 639 e-mail clicks in our data. The link-level
estimated probabilities of click-through can be compared
with a classification threshold to predict the response on any
given observation in the validation data. For example, a classification threshold of .5 means that if the estimated clicking
probability for an observation is greater than .5, the observation is classified as a click; otherwise, it is classified as a
nonclick. Similarly, the e-mail–level probabilities can be
compared with a classification threshold to predict whether
a given e-mail will elicit a click from a given user.
We use receiver operating characteristic (ROC) curves to
compare the predictive performance of our models for a
range of thresholds. An ROC analysis is used in psychology
and medical statistics for signal detection purposes (e.g.,
Egan 1975; Metz 1986; Swets 1979). It illustrates the tradeoffs between two types of errors. For any threshold that is
used for predicting observations as clicks or nonclicks, two
types of errors can occur. False negatives are errors that
occur when actual clicks are predicted as nonclicks. False
positives are errors that occur when nonresponses (i.e.,
nonclicks) are predicted as responses (i.e., clicks). The proportion of actual click-throughs that are predicted as
nonclicks is called the false negative fraction (FNF) and the
proportion of actual nonclicks that are predicted as clicks is
called the false positive fraction (FPF).
An ROC curve (also called a Lorenz diagram) is constructed by plotting the true positive fraction (1 – FPF)
against the FPF for a range of possible classification thresholds. The resulting plot is represented over a unit square.
139
Different points on the ROC curve correspond to different
classification threshold values used for prediction. The area
under the ROC curve Az provides a summary measure of the
quality of the model. A model with an ROC curve that tracks
the 45-degree line would be worthless, as it would not separate the two classes of observations at all (i.e., there would
be as many false positives as true positives). Such a curve
has Az = .5. In contrast, a perfect model would have an ROC
curve that follows the two axes and would have Az = 1. The
ROC curve, as it spans across all possible thresholds, yields
a more complete picture of predictive accuracy than measures predicated on a single classification threshold (e.g., a
hit rate using a 50% classification threshold).
Figure 1 compares the ROC curves for the three models
using the Extant data set. The top panel shows the ROC
curves based on the link-level probability estimates, and the
bottom panel shows the curves obtained from the e-mail–
level probabilities. In both panels, we observe that the two
models with heterogeneity (Model DP and Model N) have
similar predictive abilities. Moreover, the ROC curves for
the two heterogeneous models dominate the ROC curve for
the nonheterogeneous model (Model S) at all values of the
classification threshold.
Comparing the areas under the link-level ROC curves, we
find that Az = .85 for Model DP, Az = .84 for Model N, and
Az = .76 for Model S. For the e-mail–level ROC curves, Az =
.81 for Model DP, Az = .80 for Model N, and Az = .74 for
Model S. These results suggest that for our data set, Model
DP performs slightly better than Model N and that both substantially outperform a model with no heterogeneity.
For the Novel data set, the conclusions are similar. We
find that Az = .77 for Model DP, Az = .75 for Model N, and
Az = .68 for Model S. The e-mail–level ROC curves also
provide similar conclusions. Specifically, Az = .69 for
Model DP, Az = .66 for Model N, and Az = .56 for Model S.
As with the Extant data set, Model DP predicts slightly better than Model N, and both models are superior to the nonheterogeneous model. These findings reinforce the importance of modeling heterogeneity for customization
purposes.
In summary, the PsBF and the ROC analysis indicate that
the models that account for heterogeneity are preferable to
the model that ignores sources of difference in parameters.
That unobserved heterogeneity matters indicates that there
could be sufficient gains from customization. Moreover, we
find that Model DP is superior to Model N, on the basis of
PsBF, and evidences slightly better predictive performance
on the validation data.
Parameter Estimates
We now report the parameter estimates for Model DP. As
reported previously, these estimates are based on a post–
burn-in sample of 20,000 MCMC draws. Table 2 reports the
parameter estimates, which we discuss next.
Design variables. The response rates for links within text
e-mails appear to be no greater than the response rates for
links within html e-mails. As anticipated, the effect of link
order is negative, indicating that the effectiveness of links
decreases as the link appears later in the e-mail. In contrast
to our expectations, we observe no population effect for the
number of links within an e-mail. In our data, the number of
links never exceeded eight, and this may have been too few
to generate a negative effect of clutter. We find that there is
140
JOURNAL OF MARKETING RESEARCH, MAY 2003
Figure 1
ROC CURVES FOR LINK-LEVEL AND E-MAIL–LEVEL
PROBABILITIES
Notes: The dark solid curve pertains to Model DP, Model N is represented by the gray curve, and the nonheterogeneous model is depicted by
the dashed curve.
considerable design heterogeneity; users react differently to
different designs. Thus, the number of items and the link
position can influence the responses of at least some consumers in our sample. The magnitude of the e-mail heterogeneity also indicates that some e-mails evidence a greater
effect of link order than others. The finding that some
respondents are more likely to respond to text and others
respond better to html may reflect the influence of bandwidth. Because the html is bandwidth-intensive, recipients
with lesser bandwidth may prefer text-based e-mails.
Content variables. At the population level, content types
differ in their ability to elicit click-throughs. Moreover, the
standard deviations associated with the normal baseline distribution for the unobserved user random effects clearly
reveal that there is considerable user preference heterogeneity in the data. People differ in their preferences for different
content types. There is a sizable degree of heterogeneity in
content preference across e-mails. This heterogeneity presumably arises from editorial and design differences across
the e-mails (e.g., a review in one e-mail may be of more
interest than a review in another e-mail because of how it
interacts with unobserved contextual variables). When the
fixed and random effects are into account, it is clear that the
content variables play an important role in predicting clickthroughs.
Heterogeneity. The Since variable has a negative impact
on response. This suggests that the greater the duration since
the previous click, the less likely it is that a user will click
on a link within the e-mail. The precision parameters α1, α2,
and α3 associated with the Dirichlet process priors suggest
that there is greater clustering in the random effects than is
evidenced by normal population distributions. For example,
α1 ≈ 103 implies an average of 61 clusters for the users.
Similarly, α2 ≈ 115 and α3 ≈ 383 imply on average 65 clusters for the e-mails and 383 clusters for the link random
effects. This, coupled with the PsBF favoring the DP model,
indicates that the population distributions deviate from normality and justifies the need for a semiparametric approach.
It is informative to compare the different sources of heterogeneity. A variance decomposition of the random terms
in the utility function shows that user heterogeneity
accounts for 28.37%, the e-mail heterogeneity accounts for
38.99%, the link heterogeneity accounts for 1.85%, and the
residual errors account for 31.77% of the total random variation. This implies that substantial improvements in model
performance can be realized if multiple sources of heterogeneity are modeled.
CUSTOMIZED E-MAIL DESIGN
The customer- and e-mail–specific parameter vectors
obtained from the statistical model are used to forecast
potential customer reactions to proposed changes in e-mail
content and configuration. These forecasts can be used to
determine the optimal design and content of an e-mail for
each customer. Given an objective function, combinatorial
optimization is used to customize the e-mail design for each
e-mail and each user.
In e-mail marketing situations, management’s primary
goal is to maximize the expected number of click-throughs
within e-mails. A secondary objective might be to maximize
the probability that at least one link is clicked within an email. Given the objective function, design optimization
involves (1) selecting from the set of available content the
specific content (links) to be included for each person and
each e-mail and (2) configuring the e-mail layout to maximize objectives. Note that in our sample, even though the
number of links within the e-mail (Num-Items) does not
have an effect at the population level, there is considerable
E-Customization
141
Table 2
PARAMETER ESTIMATES FOR MODEL DP
Variables
Fixed
Effects
µ
95%
Probability
Interval
Standard of Base
Distribution Across
Users
Standard of Base
Distribution Across
E-Mails
Intercept
Content1
Content2
Content3
Content4
Content5
Content6
Content7
Content8
Content9
Content10
Content11
Position
Since
Num-Items
Text
α1
α2
α3
–1.47
.25
1.07
.21
.29
1.21
.79
–.67
.80
.28
.32
–1.20
–.37
–.24
.01
.29
103.32
114.25
383.07
(–2.97, –.59)
(–.48, .79)
(.57, 1.47)
(–.54, .71)
(–.56, .72)
(.21, 1.62)
(.13, 1.25)
(–2.45, .19)
(–.11, 1.43)
(–.95, .93)
(–.59, .89)
(–3.38, .27)
(–.59, –.19)
(–.59, –.19)
(–.17, .13)
(–.33, .65)
(69.37, 130.87)
(77.26, 144.88)
(319.93, 431.89)
.51
.97
.34
.65
.35
.93
.38
.49
.38
.45
.49
.54
.23
.17
.18
.45
.44
.52
.47
.71
.52
.57
.72
.52
.56
.62
1.45
.27
.28
Notes: The parentheses contain the 2.5th and the 97.5th percentiles.
heterogeneity at the user level, and therefore content selection can be important for at least some users.
Let n be the total number of content links available to be
included in a particular e-mail. For the aforementioned
objective functions, the design problem of selecting from
the n available links and then ordering the included links can
be solved in two stages. In the first stage, n linear assignment subproblems are solved. We index each of these n subproblems by k and let k ∈ {1, …, n}. In the second stage,
the solution for the first-stage subproblem that yields the
largest objective function is chosen as the optimal solution
for the original problem.
In the first stage, let us consider the kth subproblem of
assigning the n available links to k contiguous positions
within an e-mail. Let xij be a binary variable that is equal to
1 when content link i is present in ordinal position j (j = 1
for the first position and k for the lowest position) within an
e-mail, and xij = 0 otherwise. Similarly, let pij|k be the probability that the user clicks on link i if it is placed in position
j, when the total number of included items is k. When the
interest is to maximize the expected number of clickthroughs, the problem of assigning n links to k positions is
given by the following linear assignment specification:
Maximize
n
k
i =1
j=1
∑ ∑p
ij k x ij
second constraint ensures that each of the k destinations
(ordinal position) can have at most one link. Each of the n
subproblems, corresponding to k = 1 to n can be solved by
means of the Hungarian method (Foulds 1984, pp. 72–76),
which provides an efficient and simple approach. The solution to the kth subproblem yields an assignment x*k of the
content links to k contiguous positions and the optimal value
of the objective function Vk. In each problem, where k ≠ n,
n – k links are left unassigned and are therefore not included
in the e-mail.
In the second stage, the solutions to the n subproblems are
compared to determine the subproblem that yields the maximum value for the objective function. That is, the subproblem m is determined such that Vm = max{V1, …, Vn}. The
solution x*m is chosen as the solution of the original problem. This yields the optimal content and configuration of the
e-mail.
Similarly, when the interest is to maximize the probability of at least one click from an e-mail, the kth second stage
subproblem of assigning n links to k positions can be written as
Minimize
n
k
i =1
j=1
∑ ∑ log(1 − p
ij k )x ij
subject to
k
∑x
subject to
k
∑x
ij
≤ 1, for i = 1, 2, K , n
n
∑x
ij
= 1, for j = 1, 2, K , k.
i =1
The first constraint in this specification ensures that each
content link i can be in at most one of the k destinations. The
≤ 1, for i = 1, 2, K , n
ij
= 1, for j = 1, 2, K , k.
n
∑x
j=1
ij
j=1
i =1
Minimizing the objective function in this specification is the
same as maximizing the probability that at least one link is
clicked, 1 − Π ij (1 − p ij| k ) x ij, and therefore the problem is of
the linear assignment type. The optimal solution can be
142
JOURNAL OF MARKETING RESEARCH, MAY 2003
Table 3
EXTANT OPTIMIZATION RESULTS
Probability (At Least One Click)
Configurations
Original
Greedy
Ordering
Optimal
Expected Number of Clicks
Model DP
Model N
Model DP
Model N
.23
.34
.35
.36
.23
.36
.36
.39
.34
.51
.53
.55
.32
.51
.52
.57
ascertained as previously, by comparing the n first-stage
subproblems in terms of the objective function.
Optimization Results
In this section, we report the optimization results, which
are based on the validation data. In particular, we compare
and contrast the results from the optimization procedure
detailed previously (Optimal) with those obtained using two
suboptimal but simpler procedures. The first suboptimal
procedure (Ordering) uses a single linear assignment algorithm (instead of n) for each user merely to reorder the existing links within an e-mail. This procurement therefore
ignores the content-selection aspect of the optimization and
is expected to do well when clutter does not significantly
influence the probability of click-through. The second suboptimal procedure (Greedy) uses a greedy heuristic to
reorder the links. In this heuristic, links are assigned sequentially from the first position within the e-mail to the last
position. To determine which link resides in which position
for a given user, the link having the highest probability of
click-through (among the set of unassigned links) is
assigned to the highest (uppermost) remaining position.
Each algorithm yields a customized solution for each user
and e-mail, predicated on the link-level clicking likelihoods.
We report the results for each validation set described in the
“Data” section.
Extant validation data. Table 3 reports the optimization
results for the two objective functions and the three optimization procedures using the estimates from Model DP and
Model N. Columns 1 and 2 display the results when the
objective is to maximize the probability of at least one click
from an e-mail. The entries in the first two columns contain
the mean probability of at least one click across all the emails in the validation sample. Columns 3 and 4 report the
results when the objective is to maximize the expected number of click-throughs in an e-mail. The entries in these
columns give the mean of the expected number of clicks
across all the e-mails in the validation sample. Columns 1
and 3 present the results for the Model DP, and Columns 2
and 4 indicate the Model N results. As the Model DP and
Model N results are similar, we discuss only the Model DP
results.
The first row of Table 3 gives the predicted results when
the original configuration within the data is e-mailed and
therefore serves as a benchmark for assessing the performance of the various optimization approaches. For example,
Table 3 shows that mean probability of at least one click
from an e-mail is .23 when the original e-mails (as designed
by the site) are sent. The second row gives the predicted
results for the Greedy procedure, the third row gives the
results arising from the Ordering procedure, and the last row
gives the results using the Optimal two-stage procedure.
Thus, Table 3 indicates that, for Model DP, the mean probability of at least one click can increase to .36 if the optimal
e-mails are sent.
Comparing the entries in the first column, we find that
Optimal procedure can yield a 56% improvement over the
original configuration in the mean probability of at least one
click-through. In contrast, the improvement for the Ordering
procedure is 52%. Decomposing the total potential improvement into the portion arising from reordering and the portion
arising from content selection suggests that reordering
results in 92% of the total improvement, and content selection constitutes the balance. Further analysis suggests that
the Optimal algorithm improves the likelihood of at least
one click over the Ordering algorithm for 43% of the e-mails
sent out in the validation sample. A majority of the gains are
small and arise from the users who react adversely to clutter
(i.e., have a negative coefficient for the variable number of
items). Were users more averse to clutter, content selection
would matter even more.
The Greedy procedure results in a 48% increase in predicted e-mail click rates (from .23 to .34). Thus, the algorithm performs nearly as well as the Ordering algorithm.
Nonetheless, we are wary of predicting similar improvements in different data. In particular, when subjects have a
positive coefficient for the positive variable (i.e., they tend to
scroll to the bottom of the e-mail and then click on one of
the links at the bottom), the Greedy algorithm we use will
not do well.
The entries in the second column show that the Optimal
optimization procedure can yield a 62% improvement over
the original configuration in the expected number of clicks.
In contrast, the improvements for the Ordering procedure
and the Greedy algorithm are 53% and 50%, respectively.
Furthermore, for approximately 42% of the e-mails sent out
in the validation sample, Optimal made an improvement
over Ordering. Similarly, for 58% of the e-mails, Optimal
was better than Greedy, though the magnitude of improvement was small in most cases.
The Optimal procedure is better than the other two procedures and leads to improvements in response rates for emails, especially when clutter adversely affects the probability of clicking. In addition, the linear assignment problem
that forms the basis for the optimal procedure can be solved
quickly and efficiently, even for large problems. The Greedy
solution, though simpler, performs poorly for users who
have a propensity to click at the bottom of the e-mail. In
contrast, the Ordering algorithm can do relatively well in
this situation yet performs poorly when clutter matters. In
general, the Ordering procedure and the Greedy procedure
are marginally less demanding computationally. The choice
E-Customization
143
Table 4
NOVEL OPTIMIZATION RESULTS
Probability (At Least One Click)
Configurations
Original
Greedy
Ordering
Optimal
Expected Number of Clicks
Model DP
Model N
Model DP
Model N
.45
.53
.53
.54
.45
.53
.53
.54
.61
.75
.75
.75
.60
.74
.74
.75
between these depends on the balance among accuracy,
speed, and complexity desired by managers.
When applying a similar analysis using the parameters
from Model S (the nonheterogeneous model), the optimal
two-stage optimization procedure yields a predicted
improvement of only 12.5% in the mean probability of at
least one click-through and an improvement of 15.4% for
the second objective function. This relative lack of improvement predicted by the simpler model compared with the heterogeneous models arises because the simpler model has
limited flexibility in differentiating among users and emails. It therefore generates e-mail configurations that are
optimal only for the average user or e-mail.
Novel validation data. Table 4 reports the optimization
results for the Novel validation data set. Column 2 indicates
that the mean probability of at least one click increases 20%
from .45 when the original e-mails are sent to .54 when the
optimal e-mails are sent. The improvement for the Ordering
procedure is similar, 18%. Decomposing the total potential
improvement into the portion arising from reordering and
the portion arising from content selection suggests that
reordering results in 90% of the total improvement and content selection constitutes the balance.
For the second objective function, the results obtained
from the DP model estimates in the fourth column show that
the Optimal optimization procedure can yield a 23%
improvement over the original configuration in the expected
number of clicks. The improvements for the Ordering procedure and the Greedy algorithm are also 23%. The results
for the normal model are similar and are shown in the fifth
column. For both objective functions, the optimization
results for the Novel data set further suggest that any of the
three optimization approaches work equally well in these
data.
Model S yields a predicted improvement in the mean
probability of at least one click-through of 14.2% and an
improvement of 15.8% for expected clicks. However, given
that heterogeneous models predict significantly better than
the simple model with no heterogeneity, we place greater
credence on the optimization results from the heterogeneous
models.
Finally, we note that the improvement in hit rates using
the Novel data, though substantial, is not as great as the
improvement indicated by the Extant data. The difference
arises because the information regarding the clicking behavior of other users on a particular link is informative about the
targeting of links. When possible, content providers should
seek to “test market” information, as this enhances targeting
efficacy.
CONCLUSION
The advent of the Internet has enhanced the ability of
marketers to personalize communications and engender
relationships with consumers. By enabling the right content
to reach the right person at the right time, the Web can yield
substantial dividends to Web marketers and can enhance the
quality of service to consumers. However, this promise is
contingent on learning more about consumer preferences
and developing techniques that enable marketers to fulfill
these preferences. Our objective has been to facilitate that
task.
Accordingly, we describe an approach to harness the
potential afforded by the Web to determine individual-level
preferences, and then we develop an algorithm to customize
content predicated on those preferences. In the context of
targeting and customizing e-mails that herald content in a
Web magazine, we develop a customization system that uses
an MDP probit model coupled with an optimization model
to personalize communications on the Internet.
Our adaptation of the Dirichlet process model is unique
from previous implementations in that it incorporates multiple sources of heterogeneity, uses the probit framework, and
is applied to a large-scale data application. In contrast to the
normal model, the MDP model predicates a user’s choice
behavior on that of the user’s “nearest neighbors.” As such,
a comparison of the MDP and normal models provides some
insight into the additional predictive power of model-based
collaborative filtering. Our model comparison results indicate that the MDP model is preferable to the normal model
on the basis of the PsBF. The predictive performance of the
MDP model is slightly better than that of the normal model.
We leave a thorough comparison of these alternative
approaches for further research, but we note that our
approach is tailored to the problem, as opposed to the data.
By virtue of its flexibility, there may be cases in which the
Dirichlet process model substantially outperforms the normal model (i.e., nonnormal heterogeneity). The converse is
unlikely to be true, as normal models do not adjust well to
nonnormal heterogeneity.
Given that the additional programming demands inherent
in the Dirichlet process model (over that of the normal
model) are negligible, the trade-off between the approaches
is an issue of flexibility and scalability. The computational
demands of the normal model are simpler, and therefore we
recommend the normal model when scalability of the model
is a major concern. Moreover, the scalability of either model
to the demands of sending e-mails for many users requires a
careful decomposition of the overall requirements into
offline and online (i.e., real-time) components. For example,
aggregate features of the models, such as the population dis-
144
JOURNAL OF MARKETING RESEARCH, MAY 2003
tribution, can be estimated offline on the basis of a sample
of users. Moreover, the population distribution can be
updated periodically (say, weekly) as new data arrives.
When the population distribution is known, obtaining the
estimates for a particular user is not computationally intensive and can be done relatively quickly. The optimization
component is quick, as it is based on the linear assignment
problem that has quick and exact solutions.
Our approach adds to the targeting and customization literature in marketing by integrating heterogeneous choice
models with optimization techniques to personalize content.
Specifically, we describe an optimization algorithm based
on the assignment algorithm to optimize the design and content of electronic communications. We believe that such a
general approach (combining choice models with optimization models; Rossi, McCulloch, and Allenby 1996; Tellis
and Zufryden 1995) has utility beyond e-mail customization
and can be used in the design of tailored services, customdesigned catalogs, and bundling of goods.
The results of our model indicate that the design of the email is crucial in affecting click-through probabilities. For
example, we find that the order of content matters and that
there exists a great deal of heterogeneity across users in their
preferences and across links and e-mails in terms of their
effectiveness in design and content. Capitalizing on these
results, we demonstrate that design and content can indeed
be optimized. We find that response rates (expected clickthroughs) could be increased by 62% if the e-mail’s design
is customized.
Finally, we propose that our analysis be extended along
two dimensions—modeling other behaviors in the Internet
environment and using our methodological approach in
other contexts. With regard to other Internet applications, it
would be desirable to customize Web content in an effort to
increase the frequency of site visits and clicks per visit.
Similarly, our general approach could be adapted to the
design of e-commerce sites and personal agents. It is also
possible to use our underlying methodology to target advertising content. Another pressing problem centers around
optimal contact strategies for e-mail communications.
Excessive contact, or “over-touching,” can lead to unsubscribe decisions. Infrequent contact could lead to few
responses. Moreover, these effects could vary by user. With
respect to other contexts, the proposed model could be
extended beyond e-commerce models to more traditional
models of targeting purchase opportunities, such as direct
mail marketing or product customization. Our design
approach could be adapted to conjoint tasks. The conjoint
domain may prove especially promising, as affective measures can be integrated with behavioral ones to enhance the
predictive capability of the model, and the researcher may
have more latitude in the design of the stimulus set. We hope
that this analysis will encourage further research along these
dimensions.
REFERENCES
Aberdeen Group (2001), “e-Mail Marketing: Relevancy, Retention,
and ROI,” working paper, [available at http:www.aberdeen.com/
abcompany/hottopics/emailmarketing/default.htm].
Alba, Joseph, John Lynch, Barton Weitz, Chris Janiszewski,
Richard Lutz, Alan Sawyer, and Stacy Wood (1997), “Interactive
Home Shopping: Consumer, Retailer, and Manufacturer Incen-
tives to Participate in Electronic Marketplaces,” Journal of Marketing, 61 (July), 38–53.
Allenby, Greg M., N. Arora, and J.L. Ginter (1998), “On the Heterogeneity of Demand,” Journal of Marketing Research, 35
(August), 384–89.
——— and P. Rossi (1999), “Marketing Models of Consumer Heterogeneity,” Journal of Econometrics, 89 (1–2), 57–79.
Ansari, Asim, Skander Essegaier, and Rajeev Kohli (2000), “Internet Recommender Systems,” Journal of Marketing Research, 37
(August), 363–75.
———, K. Jedidi, and S. Jagpal (2000), “A Hierarchical Approach
for Modeling Heterogeneity in Structural Equation Models,”
Marketing Science, 19 (4), 328–47.
Becker, Holly, Douglas Anmuth, and Stephanie Leichter (2002),
“Online Advertising,” Internet & Media Global Equity
Research. New York: Lehman Brothers.
Blackwell, D. and J.B. MacQueen (1973), “Ferguson Distribution
via Polya Urn Schemes,” The Annals of Statistics, 1, 353–55.
Breese, John S., David Heckerman, and Carl Kadie (1998),
“Empirical Analysis of Predictive Algorithms for Collaborative
Filtering,” Technical Report MSR-TR-98-12, Microsoft
Research (October).
Bush, C.A. and S.N. MacEachern (1996), “A Semi-parametric
Bayesian Model for Randomized Block Designs,” Biometrika,
83 (2), 275–85.
Doss, H. (1994) “Bayesian Nonparametric Estimation for Incomplete Data via Successive Substitution Sampling,” The Annals of
Statistics, 22 (December), 1763–86.
Egan, J.P. (1975), Signal Detection Theory and ROC Analysis. New
York: Academic Press.
Escobar, M.D. (1994), “Estimating Normal Means with a Dirichlet
Process Prior,” Journal of the American Statistical Association,
89 (March), 268–77.
Ferguson, T.S. (1973), “A Bayesian Analysis of Some Nonparametric Problems,” The Annals of Statistics, 1, 209–30.
——— (1974), “Prior Distributions on Spaces of Probability Measures,” The Annals of Statistics, 2, 615–29.
Foulds, L.R. (1984) Combinational Optimization for Undergraduates. New York: Springer-Verlag.
Geisser, S. and W. Eddy (1979), “A Predictive Approach to Model
Selection,” Journal of the American Statistical Association, 74
(March), 153–60.
Gelfand, A.E. (1996), “Model Determination Using SamplingBased Models,” in Monte Carlo Markov Chain in Practice, W.R.
Gilks, S. Richardson, and D.J. Spiegelhalter, eds. London:
Chapman and Hall, 145–61.
Gershoff, Andrew and Patricia M. West (1998), “Using a Community of Knowledge to Build Intelligent Agents,” Marketing Letters, 9 (1), 79–91.
Hanssens, Dominique M. and Barton A. Weitz (1980), “The Effectiveness of Industrial Print Advertisements Across Product Categories,” Journal of Marketing Research, 17 (August), 294–306.
Hoque, Abeer Y. and Gerald L. Lohse (1999), “An Information
Search Cost Perspective for Designing Interfaces for Electronic
Commerce,” Journal of Marketing Research, 36 (August), 387–
94.
Houston, F.S. and Carol Scott (1984), “The Determinants of
Advertising Page Exposure,” Journal of Advertising, 13 (2), 27–
33.
MacEachern, S.N. (1994), “Estimating Normal Means with a Conjugate Style Dirichlet Process Prior,” Communications in Statistics: Simulation and Computation, 23 (3), 727–41.
Manchanda, P., A. Ansari, and S. Gupta (1999), “The Shopping
Basket: A Model for Multicategory Purchase Incidence Decisions,” Marketing Science, 18 (2), 95–114.
Mela, Carl F., K. Jedidi, and D. Bowman (1998), “The Long-Term
Impact of Promotions on Consumer Stockpiling Behavior,”
Journal of Marketing Research, 35 (May), 250–62.
E-Customization
Metz, C.E. (1986), “ROC Methodology in Radiologic Imaging,”
Investigative Radiology, 21 (9), 720–33.
Perkowitz, M. and Oren Etzioni (1997), “Adaptive Web Sites:
Automatically Learning from User Access Patterns,” in Proceedings of 6th International WWW Conference. Seattle: University of Washington.
Rossi, Peter E., Robert E. McCulloch, and Greg M. Allenby
(1996), “The Value of Purchase History Data in Target Marketing,” Marketing Science, 15 (4), 321–40.
Sarukkai, Ramesh (2000), “Link Prediction and Path Analysis
Using Markov Chains,” Computer Networks, 33 (1–6), 377–86.
Satterthwaite, R. (1999), “Survey Results: The Real Cost of
E-commerce Sites,” Gartner Group Research Note M-07-8931.
Stamford, CT: Gartner Group (May 7).
Schmittlein, D.C., D.G. Morrison, and R. Colombo (1987),
“Counting Your Customers: Who Are They and What Will They
Do Next?” Management Science, 33 (1), 1–24.
145
Shaffer, G. and Z. John Zhang (1995), “Competitive Coupon Targeting,” Marketing Science, 14 (4), 395–416.
Swets, J.A. (1979), “ROC Analysis Applied to the Evaluation of
Medical Imaging Techniques,” Investigative Radiology, 14 (2),
109–21.
Tellis, Gerard J. and Fred S. Zufryden (1995), “Tackling the
Retailer Decision Maze: Which Brands to Discount, How Much,
and When?” Marketing Science, 14 (3), 271–99.
Wedel, M. and W. Kamakura (2000), Market Segmentation: Conceptual and Methodological Foundations, 2d ed. Boston:
Kluwer.
West, M., P. Muller, and M.D. Escobar (1994), “Hierarchical Priors and Mixture Models with Applications in Regression and
Density Estimation,” in Aspects of Uncertainty: A Tribute to D.V.
Lindley, A.F.M. Smith and P.R. Freeman, eds. London: John
Wiley & Sons, 363–86.